Google announces TranslateGemma, its new open translation models to take on ChatGPT Translate

TranslateGemma

• Google launches TranslateGemma, open translation models for 55 languages
• TranslateGemma offers local deployment and open weights for developers
• Models support text and image translation, available on Kaggle and Hugging Face

Did our AI summary help?

Google has announced TranslateGemma, a new suite of open translation models designed to support multilingual translation across 55 languages. Built on the Gemma 3 architecture, TranslateGemma is aimed at researchers and developers who want efficient, locally deployable translation models rather than closed, cloud-only systems.

The announcement places TranslateGemma in direct conversation with popular AI-powered translation tools, including ChatGPT’s translation capabilities, by focusing on openness, efficiency, and local deployment.

What is TranslateGemma

TranslateGemma is released in three model sizes — 4B, 12B, and 27B parameters — covering use cases from mobile devices to cloud deployments. According to Google, the models are trained using a two-stage process that combines supervised fine-tuning with reinforcement learning, using high-quality human and synthetic translation data.

The company says the models reduce translation error rates across high-, mid-, and low-resource languages when compared with baseline Gemma models, while using fewer parameters.

How TranslateGemma takes on ChatGPT Translate

ChatGPT’s translation feature, offered through OpenAI’s models, is widely used for quick, conversational translations but operates as a closed system. TranslateGemma takes a different approach by offering open weights that developers can download, inspect, and fine-tune.

This means TranslateGemma can run locally on devices, private servers, or specific hardware setups without sending data to external servers. For enterprises and researchers working with sensitive or low-connectivity environments, this positioning directly contrasts with ChatGPT’s cloud-based translation workflow.

Language coverage and multimodal support

TranslateGemma supports 55 evaluated language pairs and has been trained on nearly 500 additional language pairs for future experimentation. Google also notes that the models retain multimodal capabilities from Gemma 3, enabling text translation within images without separate multimodal training.

This allows developers to extend translation use cases beyond plain text, including image-based content and documents.

Availability and access

TranslateGemma models are available through platforms such as Kaggle, Hugging Face, Google Colab, and Vertex AI. Google has also published a technical report detailing training methods, benchmarks, and supported languages.

With TranslateGemma, Google is positioning open translation models as a practical alternative to proprietary AI translation tools, giving developers more control over deployment, customization, and data handling.